Key Questions on Energy and AI (World Energy Outlook Special Report)

Overview

The report gathers new questions and data since the 2025 study around 11 questions: the demand outlook, efficiency, clustering, data-centre design, electricity procurement, on-site generation, financial markets, electricity prices, AI in the energy sector, physical AI and the GDP effect. Capital spending by the largest technology companies exceeded USD 400 billion in 2025 and is expected to jump another 75% in 2026; five companies alone spend more than global oil and gas production investment. IEA satellite tracking shows the capacity of AI factories has more than tripled in 18 months.

On demand: global data-centre electricity grew 17% in 2025 and consumption of AI-focused data centres grew 50%. The central projection is roughly a doubling from 485 TWh in 2025 to 950 TWh in 2030, about 3% of global demand, with AI-focused data-centre consumption tripling. Bottlenecks reduce the likelihood of aggressive near-term scenarios, though an upside case after 2030 is possible. A shortage of high-bandwidth memory is expected to last at least to end-2027. Energy per AI task is falling at least an order of magnitude a year, but video generation, reasoning and agentic tasks can use hundreds or thousands of times more energy per query than simple text.

On design and density, AI factories combine very high power density with sharp swings in IT load. Nvidia's Ampere (2020) used about 400 W chips and 32 chips per rack for about 13 kW per rack, Blackwell uses 1,000 W chips and 72 chips for 130 kW per rack, and the announced Rubin reaches 600 kW per rack, paving the way to 1 MW. AI server power density rose 11 times in 2020-2025 and rises four times more by 2027. Because GPUs run synchronously, large power swings occur from microseconds to minutes, with potential grid-stability effects. 400 V AC-centred transitional architectures will persist, but ultra-dense deployment at scale would give strong incentives for new architectures.

On procurement and on-site generation: the tech sector signed about 40% of global corporate renewable PPAs in 2025, and the SMR offtake pipeline grew from about 25 GW at end-2024 to 45 GW at end-2025 (first projects around 2030). With slow grid connections, about one-fifth of US projects are advancing with on-site natural gas, and IEA analysis finds that reliable on-site gas power for critical, variable load requires overbuilding generation 30-70% relative to demand. By 2030 about 15-27 GW of on-site gas (equivalent to 9-20 GW of demand) and 20-25 GW of data-centre batteries are expected. China holds large shares of power-electronics supply (99% of refined gallium, about 99% of LFP battery manufacturing capacity).

Other conclusions: data-centre emissions double in IEA projections to about 350 Mt in 2035, still about 2% of electricity-sector emissions. Whether data centres raise prices depends on the supply-demand balance and policy: tight systems need new investment, systems with excess supply use capital more efficiently. An AI-driven GDP boost could raise 2035 energy demand by only 1-4%. The largest barrier to AI use in the energy sector is lack of digital skills, and only 10% of global electricity consumption is covered by open electricity data policies.

Key points

1. Global data-centre electricity grew 17% in 2025 and AI-focused data-centre consumption grew 50%.
2. The central projection is 485 TWh in 2025 to 950 TWh in 2030 (about 3% of global demand); AI-focused consumption triples.
3. Capex of the largest technology companies exceeded USD 400 billion in 2025 and is expected to rise 75% in 2026.
4. Energy per AI task falls at least an order of magnitude a year; simple text queries use less than a television, but video, reasoning and agentic tasks can use hundreds or thousands of times more.
5. Nvidia examples: Ampere about 13 kW per rack, Blackwell 130 kW, announced Rubin 600 kW; by 2027 one advanced rack could peak at the demand of about 65 households.
6. AI server power density rose 11 times in 2020-2025 and rises a further four times by 2027.
7. Synchronous GPU workloads create power swings from microseconds to minutes; battery storage is critical to supply reliability.
8. Reliable on-site gas supply for critical, variable data-centre load requires overbuilding generation by 30-70% relative to demand.
9. By 2030 about 20-25 GW of batteries and 15-27 GW of on-site gas could be installed; the SMR offtake pipeline grew from 25 GW to 45 GW.
10. Data-centre emissions reach about 350 Mt in 2035, about 2% of power-sector emissions; an AI GDP boost raises 2035 energy demand by only 1-4%.

Using it in design

• For on-site generation sizing, when users choose islanded or behind-the-meter gas, apply an N+x overbuild check (30-70% above load per IEA) and model BESS for AI power swings.
• Support very high-density AI racks (hundreds of kW; 130 kW, 600 kW and beyond) in the rack catalogue.
• In power and cooling design account for synchronous AI load swings (microseconds to minutes) and evaluate UPS and battery sizing against that profile.
• Flag concentration in the power-electronics and battery supply chain (gallium, SiC, ferrite cores, LFP) as a procurement risk.
• For projects delayed by grid connection, compare hybrid solutions (on-site generation and storage, later grid connection) as an option.

Who should read it

Engineers, investors and policy makers designing AI-focused high-density data centres and deciding on power supply and on-site generation.

Document contents

1. Executive Summary
2. Introduction
3. In light of recent market trends, has the IEA adjusted its data centre demand outlook?
4. To what extent can efficiency improvements moderate AI electricity growth?
5. Are new data centres moving to locations beyond existing clusters?
6. How is AI changing data centre design?
7. How are data centre operators procuring the electricity they need?
8. Why are some data centre developers considering onsite generation?
9. What are financial markets telling us about the impact of AI on the energy sector?
10. Are data centres raising electricity prices?
11. How can AI enhance energy security and sustainability?
12. What does the rise of physical AI and robotics mean?
13. If AI drives a boom in productivity and GDP, what would it mean for energy demand?
14. Annexes

Related standards

• ISO 8528

Read the document

Publication details

Authors

International Energy Agency

Publisher

International Energy Agency

Year

2026

Document language

English

Source type

Government / public

Access

Open licence — downloadable

Licence

CC BY 4.0 — share with attribution

Licence evidence

www.iea.org

Citation